Add icefall RNN-T results

[wgb14]

We would like to add results from Icefall RNN-T recipe. The numbers are comparable to those form Conformer CTC with ngram & attention rescoring, but decoding speed is much faster.

it takes only 1 minute to decode the dev dataset

https://github.com/k2-fsa/icefall/pull/318#issuecomment-1125625918

[csukuangfj]

https://github.com/k2-fsa/icefall/pull/361 has an updated version for the results of modified beam serach

10.52/10.62 -> 10.50/10.60

[chenguoguo]

Great, great, could you adjust the numbers, and rank it accordingly?

[wgb14]

Will do more decoding experiments and update results here

[csukuangfj]

Will do more decoding experiments and update results here

You can also try --iter xx --avg xx instead of --epoch xx --avg xx.

[wgb14]

Update results here:		Dev	Test
WER	10.4	10.51

This RNN-T model outperforms other models.

[chenguoguo]

Great, merging.